Research map / Multi-agent systems and games / Human-AI and goal-conditioned
Socialized Learning: Making Each Other Better Through Multi-Agent Collaboration
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes socialized learning (SL) and MASC, a multi-agent method where a student learns from multiple teacher agents (collective collaboration) and teachers then learn new general classes via frozen-backbone classifiers with Helmholtz free energy-based selection (reciprocal altruism), evaluated on CIFAR10/CIFAR100 splits.
Problem
Enabling multiple agents, each with unique expert classes under strong data heterogeneity, to learn new general classes while directionally preserving accuracy on their original expert classes, at low communication cost.
Approach
Combines collective collaboration (student distilled from multiple teachers with cross-entropy, knowledge distillation and energy alignment losses) with reciprocal altruism (each teacher trains a new classifier on the frozen student backbone, and inference selects between expert and general classifiers by Helmholtz free energy).
Key results
- On CIFAR10 (5 agents, 2 classes each), MASC reaches 93.40 expert / 81.04 general / 87.22 average after growth, versus best baseline FedDecorr at 83.69 average.
- On CIFAR100 (4 agents, 25 classes each), MASC reaches 65.40 expert / 58.64 general / 62.02 average, versus best baseline FedSAM at 57.55 average.
- Ablation on CIFAR10: collective collaboration alone yields 63.60 average and reciprocal altruism alone 58.38; combined gives 87.22.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use MASC instead of the strongest baseline, FedDecorr | CIFAR10 learning · CIFAR10 | average accuracy after growth: +3.53 | Helped |
| use MASC instead of the strongest baseline, FedDecorr | CIFAR100 learning · CIFAR100 | average accuracy after growth: +3.47 | Helped |
| add reciprocal altruism to collective collaboration (ablation) | MASC on CIFAR10 · CIFAR10 | average accuracy: 63.60 -> 87.22 | Helped |
| add collective collaboration to reciprocal altruism (ablation) | MASC on CIFAR10 · CIFAR10 | average accuracy: 58.38 -> 87.22 | Helped |
| add reciprocal altruism to collective collaboration (ablation) | MASC on CIFAR100 · CIFAR100 | average accuracy: 61.85 -> 62.02 | Helped |
| add collective collaboration to reciprocal altruism (ablation) | MASC on CIFAR100 · CIFAR100 | average accuracy: 61.14 -> 62.02 | Helped |
Surprising findings
- Most FL methods keep expert and general class accuracy close, so they fail to prioritize expert classes directionally.
- Data-driven distillation from per-class statistics (no raw samples) improves agent performance over no distillation, per the paper's generated-data experiments.
Limitations
- Experiments limited to CIFAR10 and CIFAR100 with artificial class splits; the paper notes the lack of real-world SL datasets with many classes.
- Only two-agent theory is analyzed; the multi-agent extension is argued qualitatively.
- Expert-class accuracy in some baselines is compared against a single-agent setup that may not reflect real deployments.
Experimental setup
Datasets: CIFAR10 CIFAR100
Benchmarks: CIFAR10 CIFAR100
Metrics: expert-class accuracy general-class accuracy average accuracy
Baselines: DER EWC FOSTER iCaRL LwF MEMO PODNet WA FedAvg FedAvgM FedNova FedDecorr
Builds on: knowledge distillation Helmholtz free energy classifier selection federated averaging
Closest papers by meaning
- Socialized Coevolution: Advancing a Better World through Cross-Task Collaboration (ICML 2025 · similarity 0.90)
- Adaptive Collaboration with Humans: Metacognitive Policy Optimization for Multi-Agent LLMs with Continual Learning (ICLR 2026 · similarity 0.87)
- Artificial Generational Intelligence: Cultural Accumulation in Reinforcement Learning (NeurIPS 2024 · similarity 0.86)
- AgentVerse: Facilitating Multi-Agent Collaboration and Exploring Emergent Behaviors (ICLR 2024 · similarity 0.86)
- Peer Learning: Learning Complex Policies in Groups from Scratch via Action Recommendations (AAAI 2024 · similarity 0.86)
- Scaling Large Language Model-based Multi-Agent Collaboration (ICLR 2025 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.